Mistral's Licence Terms: Apache 2.0 Models and Research-Only Releases

“Mistral is open source” is true of some of Mistral’s weights and false of others, and the difference is the difference between shipping a product and not. The licence is a property of the individual checkpoint, not of the company.

Three licence families, not two

People usually arrive at this question with a binary in mind — open or closed. Mistral has published under at least three distinct regimes, and the middle one is the one that causes trouble because the weights are downloadable, which reads as permission.

• Apache License 2.0. A standard, permissive, OSI-approved licence. Commercial use, modification, redistribution and closed-source derivatives are all permitted, subject to keeping the notice and licence text. No field-of-use restriction, no revenue threshold, no separate agreement with Mistral.
• Mistral Research Licence (MRL). Mistral’s own text, published at mistral.ai/terms. Weights are downloadable and usable for research and non-commercial purposes. Production or commercial use requires a separate commercial licence from Mistral. This is not open source in the OSI sense and it is not a variant of Apache with extra words.
• API-only, no weights. Some models in the API catalogue have never had public weights at all. There is no licence question to resolve; there is a terms-of-service question, governed by Mistral’s platform terms.

Mistral has also used a non-production licence on at least one code model line, which sits alongside the MRL in the same category for practical purposes: downloadable, usable to evaluate and to research, not usable to serve customers without an agreement.

The reason a company publishes under two regimes at once is not inconsistency. Weights that are permissively licensed spread — they get quantised, fine-tuned, packaged into local runtimes and taught in courses, and every one of those is distribution the company did not pay for. Weights that compete directly with the hosted API are held back behind a licence that permits inspection and blocks resale. The result is a catalogue where the smaller and older models are permissive and the frontier ones are not, and the boundary moves outward over time as a generation stops being the frontier.

The Apache 2.0 releases

The Apache 2.0 side of Mistral’s catalogue is broad and it is the reason the company’s open reputation is deserved. As published on the model cards under the mistralai organisation on Hugging Face, the permissively licensed lines have included the original Mistral 7B instruct series, the Mixtral sparse mixture-of-experts models, Mistral NeMo, the Mistral Small line from its 24B generation onward, and the Pixtral 12B vision model. Each card carries a license: apache-2.0 field in its YAML front matter, which is the machine-readable form of the claim.

What Apache 2.0 gives you here is worth being precise about, because it is more than “you may use it”. You may fine-tune the weights and keep the resulting adapter or merged checkpoint private. You may serve the model to paying customers. You may embed it in a product you do not distribute source for. You must reproduce the licence and any NOTICE file with distributed copies, and you must state significant modifications. There is no share-alike obligation on your own code, which is the practical distinction from a copyleft licence.

The research-only releases

The MRL side has generally covered the frontier-tier open weights — the models where Mistral is giving away something it also sells access to. The Mistral Large weight releases and several of the larger multimodal and code checkpoints have carried it. The signal on the Hugging Face card is a license: other field with a license_name naming the Mistral Research Licence, plus a gated repository that asks you to accept terms before the files download.

The gate is the part worth noticing. A repository that asks you to agree to terms before downloading is telling you, structurally, that this is not Apache 2.0 — Apache repositories have nothing to gate on. If you clicked through that agreement six months ago on a personal account and the weights are now baked into a production image, the agreement you accepted is the one that governs.

What research-only actually restricts

The MRL text is short and worth reading in full rather than summarised, but the shape of it is: you get a non-exclusive, non-transferable, royalty-free licence for research and non-commercial purposes; derivatives inherit the same terms; production use of the model or of anything derived from it needs a separate commercial agreement. “Research purposes” is defined in the text and is narrower than “work I do at a company”.

Three consequences that catch teams out:

• Internal use is still use. A model serving an internal support tool at a commercial company is not obviously research, even though nothing is sold to a customer. If the distinction matters to you, it is a question for a lawyer and for Mistral, not for a blog post.
• Fine-tunes inherit. A LoRA adapter trained on MRL-licensed base weights is a derivative. Publishing it under a permissive licence does not launder the base.
• Distillation is addressed explicitly. Model licences in this category commonly speak to outputs and to training other models on them. Check the specific text before using an MRL-licensed model to generate a synthetic training set for a model you intend to ship.

None of this makes the research licence a trap. It is a coherent bargain — full weights, full transparency, evaluate as much as you like, pay if you productionise — and it is the same bargain several other labs have converged on. It is only a problem when somebody assumes the Apache half of the catalogue covers all of it.

Checking a specific checkpoint

The authoritative artefact is the model card in the repository you are actually pulling from, at the revision you are actually pulling. You can read the licence field without downloading the weights:

# The card's YAML front matter carries the licence fields.
curl -s https://huggingface.co/api/models/mistralai/Mistral-7B-Instruct-v0.3 \
  | jq '{id, license: .cardData.license, gated}'

A gated value that is not false is your cue that terms apply. For anything you are going to ship, record the repository id, the commit hash of the revision, and the licence string in the same place you record your other dependency licences — the same discipline pinning a dated model version applies to the hosted API.

One further trap for anyone assembling a compliance answer: the licence on the weights and the licence on the outputs are separate questions. Apache 2.0 says nothing about generated text because it is a software licence and generated text is not the software. Mistral’s position on output ownership lives in its platform terms for the API and in the model licence text for the weights, and neither is answered by the license field on the card. If your question is “can I sell what this model wrote?” rather than “can I run this model?”, you are reading the wrong document.

If you are calling Mistral models through the hosted API rather than running weights, none of the above governs you: the platform terms do, and they are the same regardless of what licence the corresponding open weights carry. Serving an MRL model through Mistral’s own API is Mistral serving it, not you.